arXiv subjects

Jun Wang

Publications and source records attributed to Jun Wang.

At least 271 records · Page 15Linked to original sources

arXiv2025

SuRe: Surprise-Driven Prioritised Replay for Continual LLM Learning

Continual learning, one's ability to adapt to a sequence of tasks without forgetting previously acquired knowledge, remains a major challenge in machine learning and a key gap between artificial and human intelligence. While regularisation and replay perform well in vision, they lag behind multi-task learning for large language models (LLMs), especially at scale with many tasks. We revisit replay and argue that two failure modes drive this gap: selection (what to rehearse) and integration (how to consolidate new knowledge). To address selection, we propose Surprise-prioritised Replay (SuRe), a simple, architecture-agnostic rule that ranks and stores the most surprising (high Negative Log-Likelihood) sequences. SuRe achieves state-of-the-art performance in the Large Number of Tasks (LNT) setting and delivers the best overall average across both Standard CL and LNT benchmarks. To address integration, we add a dual-learner design with fast and slow LoRA adapters merged via an exponential moving average (EMA), enabling rapid adaptation while stabilising long-term knowledge. Combining SuRe with the dual learner yields further gains, including improvements of up to +5 accuracy points on LNT over prior SOTA. Ablation studies confirm that our proposed method remains robust under reduced replay frequency and small buffer size, demonstrating both effectiveness and sample efficiency. Taken together, our results establish replay as a strong baseline for continual LLM fine-tuning and demonstrate that surprise-based selection and slow-weight consolidation are complementary components for mitigating catastrophic forgetting.

cs.LG↗

arXiv2025

FaithFusion: Harmonizing Reconstruction and Generation via Pixel-wise Information Gain

In controllable driving-scene reconstruction and 3D scene generation, maintaining geometric fidelity while synthesizing visually plausible appearance under large viewpoint shifts is crucial. However, effective fusion of geometry-based 3DGS and appearance-driven diffusion models faces inherent challenges, as the absence of pixel-wise, 3D-consistent editing criteria often leads to over-restoration and geometric drift. To address these issues, we introduce \textbf{FaithFusion}, a 3DGS-diffusion fusion framework driven by pixel-wise Expected Information Gain (EIG). EIG acts as a unified policy for coherent spatio-temporal synthesis: it guides diffusion as a spatial prior to refine high-uncertainty regions, while its pixel-level weighting distills the edits back into 3DGS. The resulting plug-and-play system is free from extra prior conditions and structural modifications.Extensive experiments on the Waymo dataset demonstrate that our approach attains SOTA performance across NTA-IoU, NTL-IoU, and FID, maintaining an FID of 107.47 even at 6 meters lane shift. Our code is available at https://github.com/wangyuanbiubiubiu/FaithFusion.

cs.CV↗

arXiv2025

Subjective Depth and Timescale Transformers: Learning Where and When to Compute

The rigid, uniform allocation of computation in standard Transformer (TF) architectures can limit their efficiency and scalability, particularly for large-scale models and long sequences. Addressing this, we introduce Subjective Depth Transformers (SDT) and Subjective Timescale Transformers (STT), two distinct architectures that leverage Bayesian surprise signals to dynamically route computation, learning where and when to compute within decoder-only TFs. SDT augments a decoder-only stack with alternating Decision and Dynamic layers: a Decision layer computes a full block 'posterior' and a lightweight 'prior,' while a Dynamic layer employs fixed-capacity Top-K routing based on Bayesian surprise (Expected and Unexpected Change), maintaining a static compute graph. STT extends this conditional computation to the temporal domain: a transition network predicts residual updates, forming a temporal 'change hypothesis' that informs a router to dynamically execute or bypass TF blocks for each token, managing KV-cache contributions. Both architectures exhibit the predicted shift from novelty to prediction driven gating over training, suggesting alignment with surprise based principles. While operating at reduced capacity, they offer preliminary insights into the compute-accuracy trade-offs of conditional computation. The proposed architectures establish a flexible framework for efficiency, reducing self-attention computation by 75% and KV-cache requirements by 50% within each compute skipping layer, setting a pathway for more efficient models.

cs.LG↗

arXiv2025

Stitched Polar Codes

In this paper, we introduce stitched polar codes, a novel generalization of Arıkan's regular polar codes. Our core methodology reconfigures the fundamental polarization process by stitching additional structures to enhance the reliability of less reliable information bits in the original code. This approach preserves the polar transformation structure and maintains the same encoding and decoding complexity. Thanks to the flexible configuration, stitched polar codes consistently outperform regular polar codes, effectively solving the performance degradation issue in rate-matched scenarios. Furthermore, we provide theoretical analysis on the weight spectrum and the polarization speed of stitched polar codes to prove their superiority.

cs.IT↗

arXiv2025

Instantaneous Core Loss -- Cycle-by-cycle Modeling of Power Magnetics in PWM DC-AC Converters

Nowadays, PWM excitation is one of the most common waveforms seen by magnetic components in power electronic converters. Core loss modelling approaches such as improved Generalized Steinmetz equation (iGSE) or the loss map based on composite waveform hypothesis (CWH) process the PWM excitation piecewisely, which is proven to be effective for DC DC converters. As the additional challenge in PWM DC AC converters, the fundamental-frequency sinewave component induces the "major loop loss" on top of the piecewise high-frequency segments, which however cannot be modelled on a switching cycle basis by any existing methods. To address this gap, this paper proposes a novel fundamental concept, instantaneous core loss, which is the time-domain core loss observed experimentally for the first time in history. Extending the reactive voltage cancellation concept, this work presents a method to measure the instantaneous core loss, which only contains real power loss, as a function of time. Based on measurements in evaluated soft magnetic components, it was discovered that the discharging stage exhibits higher core loss than the charging stage. A modelling approach is then proposed to break down the major loop core loss, typically an average value in the literature, into the time domain to enable cycle-by-cycle modelling of core losses in PWM converters. This work enhances the fundamental understanding of the core loss process by moving from the average model to the time-domain model.

eess.SY↗

arXiv2025

Virtual Width Networks

We introduce Virtual Width Networks (VWN), a framework that delivers the benefits of wider representations without incurring the quadratic cost of increasing the hidden size. VWN decouples representational width from backbone width, expanding the embedding space while keeping backbone compute nearly constant. In our large-scale experiment, an 8-times expansion accelerates optimization by over 2 times for next-token and 3 times for next-2-token prediction. The advantage amplifies over training as both the loss gap grows and the convergence-speedup ratio increases, showing that VWN is not only token-efficient but also increasingly effective with scale. Moreover, we identify an approximately log-linear scaling relation between virtual width and loss reduction, offering an initial empirical basis and motivation for exploring virtual-width scaling as a new dimension of large-model efficiency.

cs.LG↗

arXiv2025

Succeed or Learn Slowly: Sample Efficient Off-Policy Reinforcement Learning for Mobile App Control

Reinforcement learning (RL) using foundation models for policy approximations in multi-turn tasks remains challenging. We identify two main limitations related to sparse reward settings and policy gradient updates, based on which we formulate a key insight: updates from positive samples with high returns typically do not require policy regularisation, whereas updates from negative samples, reflecting undesirable behaviour, can harm model performance. This paper introduces Succeed or Learn Slowly (SoLS), a novel off-policy RL algorithm evaluated on mobile app control tasks. SoLS improves sample efficiency when fine-tuning foundation models for user interface navigation via a modified off-policy actor-critic approach, applying direct policy updates for positive samples and conservative, regularised updates for negative ones to prevent model degradation. We augment SoLS with Successful Transition Replay (STR), which prioritises learning from successful interactions, further improving sample efficiency. We evaluate SoLS on the AndroidWorld benchmark, where it significantly outperforms existing methods (at least 17% relative increase), including prompt-engineering and RL approaches, while requiring substantially fewer computational resources than GPT-4o-based methods with 5-60x faster inference.

cs.LG↗

arXiv2025

A General Anchor-Based Framework for Scalable Fair Clustering

Fair clustering is crucial for mitigating bias in unsupervised learning, yet existing algorithms often suffer from quadratic or super-quadratic computational complexity, rendering them impractical for large-scale datasets. To bridge this gap, we introduce the Anchor-based Fair Clustering Framework (AFCF), a novel, general, and plug-and-play framework that empowers arbitrary fair clustering algorithms with linear-time scalability. Our approach first selects a small but representative set of anchors using a novel fair sampling strategy. Then, any off-the-shelf fair clustering algorithm can be applied to this small anchor set. The core of our framework lies in a novel anchor graph construction module, where we formulate an optimization problem to propagate labels while preserving fairness. This is achieved through a carefully designed group-label joint constraint, which we prove theoretically ensures that the fairness of the final clustering on the entire dataset matches that of the anchor clustering. We solve this optimization efficiently using an ADMM-based algorithm. Extensive experiments on multiple large-scale benchmarks demonstrate that AFCF drastically accelerates state-of-the-art methods, which reduces computational time by orders of magnitude while maintaining strong clustering performance and fairness guarantees.

cs.LG↗

arXiv2025

From Experience to Strategy: Empowering LLM Agents with Trainable Graph Memory

Large Language Models (LLMs) based agents have demonstrated remarkable potential in autonomous task-solving across complex, open-ended environments. A promising approach for improving the reasoning capabilities of LLM agents is to better utilize prior experiences in guiding current decisions. However, LLMs acquire experience either through implicit memory via training, which suffers from catastrophic forgetting and limited interpretability, or explicit memory via prompting, which lacks adaptability. In this paper, we introduce a novel agent-centric, trainable, multi-layered graph memory framework and evaluate how context memory enhances the ability of LLMs to utilize parametric information. The graph abstracts raw agent trajectories into structured decision paths in a state machine and further distills them into high-level, human-interpretable strategic meta-cognition. In order to make memory adaptable, we propose a reinforcement-based weight optimization procedure that estimates the empirical utility of each meta-cognition based on reward feedback from downstream tasks. These optimized strategies are then dynamically integrated into the LLM agent's training loop through meta-cognitive prompting. Empirically, the learnable graph memory delivers robust generalization, improves LLM agents' strategic reasoning performance, and provides consistent benefits during Reinforcement Learning (RL) training.

cs.CL↗

arXiv2025

Versatile and Risk-Sensitive Cardiac Diagnosis via Graph-Based ECG Signal Representation

Despite the rapid advancements of electrocardiogram (ECG) signal diagnosis and analysis methods through deep learning, two major hurdles still limit their clinical adoption: the lack of versatility in processing ECG signals with diverse configurations, and the inadequate detection of risk signals due to sample imbalances. Addressing these challenges, we introduce VersAtile and Risk-Sensitive cardiac diagnosis (VARS), an innovative approach that employs a graph-based representation to uniformly model heterogeneous ECG signals. VARS stands out by transforming ECG signals into versatile graph structures that capture critical diagnostic features, irrespective of signal diversity in the lead count, sampling frequency, and duration. This graph-centric formulation also enhances diagnostic sensitivity, enabling precise localization and identification of abnormal ECG patterns that often elude standard analysis methods. To facilitate representation transformation, our approach integrates denoising reconstruction with contrastive learning to preserve raw ECG information while highlighting pathognomonic patterns. We rigorously evaluate the efficacy of VARS on three distinct ECG datasets, encompassing a range of structural variations. The results demonstrate that VARS not only consistently surpasses existing state-of-the-art models across all these datasets but also exhibits substantial improvement in identifying risk signals. Additionally, VARS offers interpretability by pinpointing the exact waveforms that lead to specific model outputs, thereby assisting clinicians in making informed decisions. These findings suggest that our VARS will likely emerge as an invaluable tool for comprehensive cardiac health assessment.

cs.AI↗

arXiv2025

A Lightweight Complex-Valued Deformable CNN for High-Quality Computer-Generated Holography

Holographic displays have significant potential in virtual reality and augmented reality owing to their ability to provide all the depth cues. Deep learning-based methods play an important role in computer-generated holography (CGH). During the diffraction process, each pixel exerts an influence on the reconstructed image. However, previous works face challenges in capturing sufficient information to accurately model this process, primarily due to the inadequacy of their effective receptive field (ERF). Here, we designed complex-valued deformable convolution for integration into network, enabling dynamic adjustment of the convolution kernel's shape to increase flexibility of ERF for better feature extraction. This approach allows us to utilize a single model while achieving state-of-the-art performance in both simulated and optical experiment reconstructions, surpassing existing open-source models. Specifically, our method has a peak signal-to-noise ratio that is 2.04 dB, 5.31 dB, and 9.71 dB higher than that of CCNN-CGH, HoloNet, and Holo-encoder, respectively, when the resolution is 1920$\times$1072. The number of parameters of our model is only about one-eighth of that of CCNN-CGH.

physics.optics↗

arXiv2025

CoFineLLM: Conformal Finetuning of LLMs for Language-Instructed Robot Planning

Large Language Models (LLMs) have recently emerged as planners for language-instructed agents, generating sequences of actions to accomplish natural language tasks. However, their reliability remains a challenge, especially in long-horizon tasks, since they often produce overconfident yet wrong outputs. Conformal Prediction (CP) has been leveraged to address this issue by wrapping LLM outputs into prediction sets that contain the correct action with a user-defined confidence. When the prediction set is a singleton, the planner executes that action; otherwise, it requests help from a user. This has led to LLM-based planners that can ensure plan correctness with a user-defined probability. However, as LLMs are trained in an uncertainty-agnostic manner, without awareness of prediction sets, they tend to produce unnecessarily large sets, particularly at higher confidence levels, resulting in frequent human interventions limiting autonomous deployment. To address this, we introduce CoFineLLM (Conformal Finetuning for LLMs), the first CP-aware finetuning framework for LLM-based planners that explicitly reduces prediction-set size and, in turn, the need for user interventions. We evaluate our approach on multiple language-instructed robot planning problems and show consistent improvements over uncertainty-aware and uncertainty-agnostic finetuning baselines in terms of prediction-set size, and help rates. Finally, we demonstrate robustness of our method to out-of-distribution scenarios in hardware experiments.

cs.RO↗

arXiv2025

A Principle of Targeted Intervention for Multi-Agent Reinforcement Learning

Steering cooperative multi-agent reinforcement learning (MARL) towards desired outcomes is challenging, particularly when the global guidance from a human on the whole multi-agent system is impractical in a large-scale MARL. On the other hand, designing external mechanisms (e.g., intrinsic rewards and human feedback) to coordinate agents mostly relies on empirical studies, lacking a easy-to-use research tool. In this work, we employ multi-agent influence diagrams (MAIDs) as a graphical framework to address the above issues. First, we introduce the concept of MARL interaction paradigms (orthogonal to MARL learning paradigms), using MAIDs to analyze and visualize both unguided self-organization and global guidance mechanisms in MARL. Then, we design a new MARL interaction paradigm, referred to as the targeted intervention paradigm that is applied to only a single targeted agent, so the problem of global guidance can be mitigated. In implementation, we introduce a causal inference technique, referred to as Pre-Strategy Intervention (PSI), to realize the targeted intervention paradigm. Since MAIDs can be regarded as a special class of causal diagrams, a composite desired outcome that integrates the primary task goal and an additional desired outcome can be achieved by maximizing the corresponding causal effect through the PSI. Moreover, the bundled relevance graph analysis of MAIDs provides a tool to identify whether an MARL learning paradigm is workable under the design of an MARL interaction paradigm. In experiments, we demonstrate the effectiveness of our proposed targeted intervention, and verify the result of relevance graph analysis.

cs.AI↗

arXiv2025

UI-Evol: Automatic Knowledge Evolving for Computer Use Agents

External knowledge has played a crucial role in the recent development of computer use agents. We identify a critical knowledge-execution gap: retrieved knowledge often fails to translate into effective real-world task execution. Our analysis shows even 90% correct knowledge yields only 41% execution success rate. To bridge this gap, we propose UI-Evol, a plug-and-play module for autonomous GUI knowledge evolution. UI-Evol consists of two stages: a Retrace Stage that extracts faithful objective action sequences from actual agent-environment interactions, and a Critique Stage that refines existing knowledge by comparing these sequences against external references. We conduct comprehensive experiments on the OSWorld benchmark with the state-of-the-art Agent S2. Our results demonstrate that UI-Evol not only significantly boosts task performance but also addresses a previously overlooked issue of high behavioral standard deviation in computer use agents, leading to superior performance on computer use tasks and substantially improved agent reliability.

cs.HC↗

arXiv2025

When, What, and How: Rethinking Retrieval-Enhanced Speculative Decoding

Speculative decoding (SD) has emerged as an effective technique to accelerate large language model (LLM) inference without compromising output quality. However, the achievable speedup largely depends on the effectiveness of the drafting model. While model-based methods like EAGLE-2 are accurate but costly, retrieval-enhanced methods like SAM-Decoding rely on heuristic switching strategies that often trigger unnecessary retrievals. To address this, we propose ReSpec (\textbf{Re}trieval-enhanced \textbf{Spe}culative Decoding), a novel framework that transforms heuristic drafter switching into adaptive decision-making. ReSpec features three core innovations: 1) An \textbf{entropy-guided adaptive trigger} quantifies contextual predictability to initiate retrieval only when uncertainty is low, avoiding costly low-quality speculations. 2) A \textbf{feedback-driven candidate selection} leverages historical feedback to organize multiple high-quality candidates for parallel verification, maximizing retrieval utility. 3) A source-aware \textbf{relaxed verification strategy} applies strict checks to model-generated drafts while using a relaxed verification for retrieved drafts, achieving a better balance between accuracy and efficiency. Extensive experiments on Spec-Bench demonstrate that ReSpec achieves state-of-the-art acceleration,outperforming EAGLE-2 and SAM-Decoding by over $33\%$ and $25\%$, respectively, while maintaining output quality.

cs.CL↗

arXiv2025

A Comprehensive Survey on Process-Oriented Automatic Text Summarization with Exploration of LLM-Based Methods

Automatic Text Summarization (ATS), utilizing Natural Language Processing (NLP) algorithms, aims to create concise and accurate summaries, thereby significantly reducing the human effort required in processing large volumes of text. ATS has drawn considerable interest in both academic and industrial circles. Many studies have been conducted in the past to survey ATS methods; however, they generally lack practicality for real-world implementations, as they often categorize previous methods from a theoretical standpoint. Moreover, the advent of Large Language Models (LLMs) has altered conventional ATS methods. In this survey, we aim to 1) provide a comprehensive overview of ATS from a ``Process-Oriented Schema'' perspective, which is best aligned with real-world implementations; 2) comprehensively review the latest LLM-based ATS works; and 3) deliver an up-to-date survey of ATS, bridging the two-year gap in the literature. To the best of our knowledge, this is the first survey to specifically investigate LLM-based ATS methods.

cs.AI↗

arXiv2025

SparsePO: Controlling Preference Alignment of LLMs via Sparse Token Masks

Direct alignment algorithms have proven an effective step for aligning language models to human-desired behaviors. Current variants of the Direct Preference Optimization objective have focused on a strict setting where all tokens are contributing signals of KL divergence and rewards to the loss function. However, human preference is not affected equally by each word in a sequence but is often dependent on specific words or phrases, e.g. existence of toxic terms leads to non-preferred responses. Based on this observation, we argue that not all tokens should be weighted equally during PO and propose a flexible objective termed SparsePO, that aims to automatically learn to weight the KL divergence and reward corresponding to each token during PO training. We propose two different variants of weight-masks that can either be derived from the reference model itself or learned on the fly. Notably, our method induces sparsity in the learned masks, allowing the model to learn how to best balance reward and KL divergence contributions at the token level, learning an optimal level of mask sparsity. Extensive experiments illustrate the effectiveness of our approach at aligning to preference proxies, including sentiment control, helpfulness and harmlessness, and summary quality. Our method obtains +10% and +3% win rate points in summarization and dialogue scenarios, respectively, without compromising model reasoning or the relevancy and faithfulness of the summary response.

cs.CL↗

arXiv2025

Classifier Enhancement Using Extended Context and Domain Experts for Semantic Segmentation

Prevalent semantic segmentation methods generally adopt a vanilla classifier to categorize each pixel into specific classes. Although such a classifier learns global information from the training data, this information is represented by a set of fixed parameters (weights and biases). However, each image has a different class distribution, which prevents the classifier from addressing the unique characteristics of individual images. At the dataset level, class imbalance leads to segmentation results being biased towards majority classes, limiting the model's effectiveness in identifying and segmenting minority class regions. In this paper, we propose an Extended Context-Aware Classifier (ECAC) that dynamically adjusts the classifier using global (dataset-level) and local (image-level) contextual information. Specifically, we leverage a memory bank to learn dataset-level contextual information of each class, incorporating the class-specific contextual information from the current image to improve the classifier for precise pixel labeling. Additionally, a teacher-student network paradigm is adopted, where the domain expert (teacher network) dynamically adjusts contextual information with ground truth and transfers knowledge to the student network. Comprehensive experiments illustrate that the proposed ECAC can achieve state-of-the-art performance across several datasets, including ADE20K, COCO-Stuff10K, and Pascal-Context.

cs.CV↗